P8 observability with a CUDA event

A timing-disabled CUDA event reduces per-launch synchronization cost relative to cudaStreamSynchronize while preserving ~20 ms observation.

Claim statusNO END-TO-END MINING ADVANTAGE DEMONSTRATED

Diagnostic work, exact algebra and local capabilities are not treated as end-to-end mining advantage.

Date2026-08-16

HardwareNVIDIA GeForce RTX 5080 · exact CUDA SHA-256d verifier

Run scopePreregistered paired campaign with sealed validation and audit

ReproducibilityRebuild the three sealed binaries and repeat equal-work balanced measurements; do not select from the point estimate.

01 / Question & hypothesis

What was tested?

A timing-disabled CUDA event reduces per-launch synchronization cost relative to cudaStreamSynchronize while preserving ~20 ms observation.

02 / Scientific basis

Why the test is meaningful

CUDA events and stream synchronization expose different host/runtime paths; any claimed saving must reproduce across headers while retaining the no-sync baseline.

R_ES=throughput(event-sync)/throughput(stream-sync)retention=throughput(event-sync)/throughput(P8-S0)selection requires consistent R_ES>1

FIGURE / RESULT READINGVisual reading of the published metrics and gates for EXP-083; it summarizes the registered result, not a mining advantage.

03 / Method

How it was tested

Compare P8-S0, P8 stream-sync and P8 timing-disabled event-sync over four new headers and 24 runs with unchanged SHA-256d arithmetic.

04 / Observed result

What happened

Event / stream1.000022×

P8 retention0.999572

Positive headers1 / 4

Event-sync versus stream-sync was 1.000022×, CI95 0.999696–1.000347, favorable on only 1/4 headers. It retained 0.999572 of P8-S0 and observed the host every 0.020270 s.

05 / Validation

Exactness and statistical controls

All three binaries found B32, used 44 registers, had zero spills and matched compiled factors.

06 / Interpretation

What the result means

Event synchronization does not reproducibly reduce the cost. The next mechanism must synchronize less often while keeping the ~20 ms window.

Limitations

Only timing-disabled CUDA events were tested.
The interval is consistent with no difference.
No new energy campaign is justified for an unselected mechanism.

07 / Reproduction

Evidence trail

Rebuild the three sealed binaries and repeat equal-work balanced measurements; do not select from the point estimate.

Canonical variants

CANONICAL-EXP-159PREREGISTERED-CAMPAIGNSEALED-AUDIT

Source: internally audited canonical reports. Local filesystem structure, private headers and operational identifiers are excluded from publication.